An Object-Based Bayesian Framework for Top-Down Visual Attention